Computer Recognition of Two Overlapping Parts Using a Single Camera

The combination of computer vision to recognize overlapping objects and computer-based robots to acquire, orient, and transport these objects will revolutionize manufacturing technology. The bin-picking application, where robots transport randomly scattered parts from a bin and place them on a conveyor with the proper orientation, will significantly improve and economize many manufacturing processes. Equipping the robot with "sight" and the ability to recognize overlapping objects is a first step in solving the bin-picking problem. Recently, many image processing techniques have been applied to the recognition of non-overlapping objects. 1,2 The most successful of these use histogram techniques for segmenting objects from the background and then finding local features such as holes, curves, and corners;3 gradient analysis and Laplacian derivatives for locating edges; 4 and chain-encoding for representing edges and boundaries." Engineers have also employed line-thinning , line-merging, tracking, smoothing , and filtering techniques for image enhancement and sharpening;' and Opinions, findings, and conclusions or recommnenda-dons expressed in this article are those of the authors. maximal cliques,7 template matching, edge-cues, 8 and local feature cues9 for recognizing the object, its position and orientation. (Several techniques are described in the glossary following this article.) A system that determines the position and orientation of non-overlapping objects, and passes that information to a PUMA robot arm for acquisition and transporation, has already been developed at the University of Michigan's Robotics Research Laboratory. The basic method employed by this system is quite similar to the SRI vision module. '0 It compares previously taught parameters (or features) such as the area, perimeter, ratio of perimeter squared divided by area, and a radial template with similar parameters from the image calculated in real time. This system and others like it have been successful in recognizing discrete parts; however, they show severe limitations when attempting to recognize overlapping parts. The major problem lies in segmenting the image II-distinguishing two overlap